Understanding what types of cells are present in a tissue sample is key to unlocking how that tissue functions in health and disease. One way scientists estimate this information is through computational deconvolution, a method that uses RNA sequencing data to predict the proportions of different cell types within a mixed tissue sample.

Researchers at the University of North Carolina at Charlotte have investigated how well different deconvolution methods hold up under various conditions. Their team compared two main types of methods: reference-based, which rely on known gene expression profiles from specific cell types, and reference-free, which work without that prior knowledge.

Study design

(a) A cell-level dataset is used to generated pseudo-bulk tissues. (b) Ideal mixing to generate pseudo-bulk tissue RNA-seq data. (c) Deconvolution for cell proportions. (d) Performance evaluation.

To test these methods, the researchers created synthetic bulk RNA sequencing data from real single-cell datasets representing four types of tissue. They then altered the cell compositions and gene expression patterns to simulate different levels of complexity and noise. By comparing each method’s predictions to the known truth, they measured accuracy using several statistical approaches.

The results show that reference-based methods work best when high-quality reference data is available, while reference-free methods are more adaptable in situations where such data is missing. The study also highlights how variability in cell makeup and gene expression can significantly affect the performance of deconvolution algorithms.

These findings help scientists choose the most appropriate tool for their data and push the field forward by identifying what makes a method more or less reliable.